Glossary¶

The terms these pages use, each with the page that explains it.

Catalog: The set of ontologies OntoCast knows: those loaded at startup, uploaded to /ontologies, or created by earlier runs. Every ontology context is taken from it. See Ontologies and facts.
Completion pass: An optional extra call, after the critic, that asks the model for measurements the render missed. It only adds facts. See How OntoCast works.
Consolidation: An optional LLM call that refines a document's ontology against an excerpt of the whole document, after the units' changes are merged (ENABLE_ONTOLOGY_CONSOLIDATION). See How OntoCast works.
Content unit: One part of a document, cut along its sections, that is extracted on its own and in parallel with the others. See How OntoCast works.
Critic: The LLM call that reviews a render and names the statements to fix. OntoCast applies the fixes as a patch without another call. See Validation layers.
Entity disambiguation: Deciding which entities from different content units are the same thing, and merging them. See Entity disambiguation.
Facts: The graph of what a document states: the entities it mentions and their relations and values, in the terms of the ontology. See Ontologies and facts.
Facts namespace (cd:): https://growgraph.dev/facts/, where every entity extracted from text is named. The ontology's own namespaces are never extended by facts. See Ontologies and facts.
GraphUpdate (patch): What the model returns instead of a whole graph: the triples to insert and the triples to delete. See Ontologies and facts.
Ontology: A graph of classes, properties and declared individuals, identified by its IRI and stored in versions. See Ontologies and facts.
Ontology context: The slice of the catalog a content unit is shown: one chosen ontology, the terms vector retrieval finds, or one fixed ontology (ONTOLOGY_CONTEXT_MODE). See Choosing ontology context.
Provenance: The link from each fact to the content unit it came from, written with RDF 1.2 reifiers. See Ontologies and facts.
Render: The LLM call that extracts an ontology patch or a unit's facts. See How OntoCast works.
Render mode: Which blocks a run executes: ontology_and_facts, ontology or facts (RENDER_MODE). See How OntoCast works.
Run manifest: The JSON file ontocast process writes beside each facts file (<name>.run.json): the model, the settings and what the run cost. See Reading run telemetry.
Section label: The section a content unit came from, such as methods or results, used to select what to extract. See Structured documents.
Section schema: The set of section labels for one kind of document, such as academic or financial. See Structured documents.
SHACL shape: A constraint the extracted facts are validated against, loaded from FACTS_SHAPES_DIR or uploaded to /shapes. See Validation and SHACL.
Tenant and project: Request parameters that partition the triple store and the vector store, so that several users or projects share one server without seeing each other's graphs. See Tenancy.
Term sheet: A compact form of the ontology context in facts prompts: one line per term instead of a serialized graph (ONTOLOGY_CHAPTER_FORMAT). See Performance tuning.
Unit pipeline: The per-unit loops run on a single piece of text, with no conversion, chunking or merging: POST /process_unit or run_unit_pipeline. See Embedding OntoCast in your agent.
User instruction: Your own guidance, added to the prompts of one request (facts_user_instruction, ontology_user_instruction). See Writing user instructions.
Vector retrieval: Choosing a unit's ontology context by searching an index of the whole catalog (LanceDB or Qdrant) for the terms the unit needs. See Retrieval.